Abstract Details
Abstracts
Author: Andrew M. Irvin
Requested Type: Consider for Invited
Submitted: 2026-02-27 12:38:30
Co-authors: L. Casali, S. De Pascuale
Contact Info:
University of Tennessee - Knoxville
863 Neyland Dr
Knoxville, TN 37916
United States
Abstract Text:
In advanced tokamak scenarios, electron cyclotron heating and current drive plays a crucial role, enabling precise control of the plasma current profile and the suppression of neoclassical tearing modes (NTMs) via localized current drive. Raytracing codes such as TORAY can compute EC H/CD profiles quickly yet are still too slow for real-time applications and remain cumbersome for large parameter scans. Thus, it is desirable to develop a fast and accurate surrogate model using ML/AI which can compute these profiles on a millisecond timescale while preserving fidelity, to aid in the control and design of present and future advanced tokamaks. To develop a viable surrogate model which can be used across multiple devices and scenarios, training data across a broad parameter space encompassing the extent of existing and planned tokamaks in required. Sampling a sufficiently broad highly dimensional parameter space to encompass both present tokamaks and the scope of future devices using independent random sampling of inputs is inadequate to produce a large set of physically reasonable inputs for simulations used to train the surrogate model. Thus, we employ a physically constrained quasi-random sampling scheme based on established tokamak operational constraints, which ensures most cases provide useful data on which to train a ML/AI surrogate model. Utilizing this technique, we build upon previous efforts [Irvin et al. (2026) Fus. Sci. and Tech. 10.1080/15361055.2025.2476829 and Irvin et al. (2026) Fus. Eng. Des. (submitted)] to train a gaussian process regression (GPR) surrogate model to predict EC H/CD profiles in both present and future tokamaks using data from self-consistent integrated modeling between equilibrium, transport and raytracing. The surrogate model accurately reproduces radial profiles for off-axis EC H/CD across a broad range of plasma, device and EC system parameters including both present and future tokamaks on a sufficiently accelerated timescale.
Characterization: 6.0
Comments: